Improved Pectoral Muscle Segmentation in Mammograms through Regression-Based Deep Learning and Knowledge Distillation

Pectoral muscle (PM) segmentation is an important step for improving the accuracy and efficiency of breast cancer screening in digital mammography. In recent years, image-to-image (I2I) deep learning (DL) methods have achieved state-of-the-art performance for automated PM segmentation by representing the PM region as a binary mask. This paper introduces a new curve regression approach by representing the PM boundary as a vector of connected points that lie on the curve separating the PM from surrounding breast tissue. This low-dimensional PM representation is used to introduce a concept of knowledge distillation (KD), which exploits an ensemble of teachers to perform loss weighting based on maximum likelihood (ML) estimation. Experiments with in-house mammography data show that DL based curve regression outperforms a reference I2I DL method (U-net) for PM segmentation. Further, application of the proposed KD concept achieves higher segmentation accuracy with only 16% of parameters and 23% of inference time compared to the U-net.